Application of machine-learning tools to extract patterns in long-term DOC monitoring data: an integrated, multi-scale approach

Collaborative and transdisciplinary research at long-term monitoring (LTM) sites maintained under the USGS Hydro-Climatic Data Network and NSF Long Term Ecological Research network has generated a wide variety of stream water quality data that can be used to test hypotheses on ecosystem response to environmental change. The sheer volume of data, and the considerable variance in data types and temporal and spatial resolution, present a challenge for data analysis using traditional statistical methods. Machine-learning tools have evolved to identify patterns in “big data” and are increasingly used for dimension reduction, feature extraction, and trend identification. In this work, we used nonparametric, machine-learning algorithms to examine geologic, topographic, hydrologic and land cover variables for a subset of the LTM sites to define clusters of sites with similar biogeophysical conditions and to elicit patterns in dissolved organic carbon (DOC) dynamics across clusters. Catchment characteristics and time-series data for DOC and related constituents (e.g., sulfate, nitrate, calcium, pH, temperature) were mined from publically-available repositories (e.g., CAMELS: Catchment Attributes and MEteorology for Large-sample Studies; USGS National Water Information System). By pre-processing the LTM sites by cluster, spatial and temporal patterns and key drivers of stream water DOC flux were identified at regional (>100 km) and annual scales. We identified select catchment sites representative of these clusters, where greater temporal and spatial resolution of time-series records are available, to carry out process-based modeling of DOC dynamics aided by bench-scale soil experiments. This integrated, multi-scale approach, is being applied to test hypotheses concerning the origin of increasing DOC flux from forested headwater streams.